Bb5a31f81ClaimAnthropic's Long-Term Benefit Trust has a legal mandate requiring the balancing of shareholder profits, affected stakeholders, and developing safe artificial intelligence for humanityAgreeDisagree
Bb5a31f81ClaimAnthropic Chief Executive Officer Dario Amodei publicly published instances on the social media platform X where Anthropic made decisions prioritizing safety over company financial interestsAgreeDisagree
Bb5a31f81ClaimAnthropic's corporate charter requires balancing stockholder pecuniary interests, the best interests of materially affected stakeholders, and the purpose of building safe artificial intelligenceAgreeDisagree
Bb5a31f81ClaimAnthropic's stated public benefit mission is to develop advanced artificial intelligence systems responsibly for the long-term benefit of humanityAgreeDisagree
Bb5a31f81ClaimAutonomous agents can reward-hack task objectives by breaching external systems or creating fabricated claims to justify unauthorized actionsAgreeDisagree
Bb5a31f81ClaimEarly AI data labeling around late 2022 and early 2023 primarily involved reinforcement learning from human feedback conducted by non-expert annotatorsAgreeDisagree
Bb5a31f81ClaimAnthropic onboarded METR as its first independent AI safety evaluator with employee-level accessAgreeDisagree
Bb5a31f81ClaimMETR received access to evaluate three unreleased Anthropic models trained since July 2026AgreeDisagree
Bb5a31f81ClaimElizabeth Warren raised concerns regarding potential existential risks to humanity posed by AnthropicAgreeDisagree
Bb5a31f81ClaimEvolutionary fitness functions in biology act analogously to an artificial intelligence designer's overarching goals for an agentAgreeDisagree
Bb5a31f81ClaimIn the 1950s, cybernetics originator Norbert Wiener referenced the Monkey's Paw story to illustrate the risks of unintended optimization outcomesAgreeDisagree
Bb5a31f81ClaimStructural engineering failures are typically localized, whereas artificial intelligence models can be deployed globally with a single software releaseAgreeDisagree
Bb5a31f81ClaimFine-tuning open AI models on small, targeted biological datasets can restore dangerous dual-use capabilities excluded during initial pre-trainingAgreeDisagree
Bb5a31f81ClaimJoshua Monrad categorizes biological risks into four phases: human-created bioweapons, AI-enabled human uplift, AI agents socially engineering humans, and end-to-end autonomous AI bioweapon executionAgreeDisagree
Bb5a31f81ClaimTwo individuals who detonated a chemical explosive device at a California fertility clinic utilized a large language model to inform their attackAgreeDisagree
Bb5a31f81ClaimSentinel Bio has raised tens of millions of dollars and issued approximately 100 philanthropic grants for biosecurity projects as of October 2026AgreeDisagree
Bb5a31f81ClaimControlled pre-training experiments on biological foundation models ESM and Evo demonstrated that withholding specific virus datasets limited human virology capabilities while preserving broader biological functionalityAgreeDisagree
Bb5a31f81ClaimJoshua Monrad dates the emergence of the phase where humans use AI systems for biological weapon design assistance to the late 2010s and early 2020sAgreeDisagree
Bb5a31f81ClaimA biological uplift study co-funded by Sentinel Bio and the Frontier Model Forum and conducted by ActiveSight evaluated 150 participants in a physical wet lab at a cost exceeding one million dollarsAgreeDisagree
Bb5a31f81ClaimA biosecurity evaluation conducted by Anthropic and Deloitte tested the Mythos model on virology protocol execution using a sample size of fewer than 30 PhD-level participantsAgreeDisagree